Drugprices in Progress

Price Sensitivity of Demand for Prescription Drugs:
Exploiting a Regression Kink Design

Marianne Simonsen
Lars Skipper
Niels Skipper
School of Economics
Aarhus School of
School of Economics
and Management
Business
and Management
Aarhus University
Aarhus University
Aarhus University
January 2011
Abstract:
This paper investigates price sensitivity of demand for prescription drugs, using drug purchase
records for a 20 % random sample of the Danish population. We identify price responsiveness by
exploiting exogenous variation in prices caused by kinked reimbursement schemes and implement a
regression kink design. Thus, within a unifying framework we uncover price sensitivity for different
subpopulations and types of drugs. The results suggest low average price responsiveness with
corresponding price elasticities ranging from -0.09 to -0.25, implying that demand is inelastic.
However, individuals with lower education and income are more responsive to the price. Also,
essential drugs that prevent deterioration in health and prolong life have lower associated average
price sensitivity.
Key words: Prescription drugs, price, reimbursement schemes, regression kink design
JEL codes: I11, I18
Acknowledgements: We greatly appreciate valuable comments from Sren Leth-Petersen, Helena

Skyt Nielsen, Emilia Simeonova, Michael Svarer, and Torben Srensen. The paper has benefited
from comments received at the 2007 iHEA conference, the 1st meeting of the Danish
Microeconometric Network 2009, and at seminars at the Stockholm School of Economics, the
Swedish Institute for Social Research, University of Stockholm, IZA, Bonn and Carlos III, Madrid.
The usual disclaimer applies.
1. Introduction
During 1998-2007, real pharmaceutical spending within the OECD went up almost 50 %, reaching
more than USD 650 billion in 2007. Furthermore, drug spending (prescription and non-prescription)
amounted to an impressive 15 % of total health spending with a US growth that was more than
twice that of total health expenditures; see OECD (2009). Spending on pharmaceuticals is foreseen
to increase even further in the future, putting severe pressure on health care budgets. To control
such spending, there is increased interest in designing optimal health insurance schemes. Necessary
inputs into the construction of health insurance policies are reliable estimates of price sensitivity of
demand, yet the empirical evidence in this area is limited, and the existing estimates vary
considerably. Furthermore, different subgroups in the population do not invest the same in their
own health and do not face the same budget constraints. To the extent that subgroups will also react
differently to prices of prescription drugs, subgroup specific estimates are crucial.
Clearly, the task of uncovering price sensitivity of demand for drugs is complicated by the standard
selection problem: individuals who purchase prescription drugs do not constitute a random part of
the population; presumably they have a higher willingness to pay than individuals who refrain from
buying. Moreover, conditional on purchasing, the price people face is non-random. Even if studies
attempt to deal with these essential identification problems, they are often not able to distinguish
spending on one type of prescription drug from other types, or even from different types of health
care usage such as hospital admittance. In other cases, the type of drug is known, but the price is
not. This is typically solved by a combination of substantial assumptions and an imputation
strategy. Finally, identification strategies often restrict the results to hold only for a specific
subgroup of the overall population.
The perhaps most widely known study within this field is the RAND Health Insurance Experiment
(HIE), which ran from the late 1970s to the start of the 1980s. The unique feature of the HIE was
that insurance plans for non-aged individuals (61 or below) were randomly assigned in six different
locations across the US. The insurance plans differed in the size of the deductible amount, coinsurance rates and full stop-loss; an annual limit on out-of-pocket expenditures. In this sense, the
study had the flavor of a randomized experiment, which has led many to consider it the gold
standard in the literature. In the study, the average price elasticity of overall health care demand,
including prescription drugs, hospital utilization and other health care services, was estimated to be
-0.20, see Manning et al. (1987) and Newhouse (1993).
While well executed and evaluated, an important limitation of the RAND HIE was that it did not
consider the elderly population, which accounts for a large share of medical expenditures.
Recognizing this point, Contoyannis et al. (2005) estimate the price elasticity for prescription drugs
in the presence of a nonlinear price schedule for the elderly people (age 65 or over) enrolled in the
Quebec Public Pharmacare program in Canada, exploiting time variation in cost-sharing. Their
overall finding is a price elasticity ranging from -0.12 to -0.16. Similarly, Chandra et al. (2010)
study price responsiveness for people enrolled in California Public Employees Retirement System
(CalPERS) with focus on the elderly population. In 2001 co-payments went up for the fraction of
enrollees under Preferred Provider Organizations (PPO), and in 2002 co-payments were increased
for the fraction that received care through HMOs. Applying a difference-in-difference framework,
the authors estimate drug utilization elasticities with respect to patient cost. Resulting elasticities are
in line with the RAND HIE.
Acknowledging the fact that price elasticities for prescription drugs are likely to be just as
heterogeneous across types of drugs as price elasticities for other types of products, Goldman et al.
(2004) and Landsman et al. (2005) study price responsiveness for different types of therapeutic
groups. Using regression type analyses on pharmacy claims data combined with cross-sectional
variation in health plan benefits designs, both studies find that drugs used for chronic conditions are
less price sensitive (-0.1 to -0.2) than drugs used for more acute conditions (-0.3 to -0.6).
Furthermore, Tamblyn et al. (2001) consider both the elderly population and welfare participants in
Quebec, Canada and find that demand for essential drugs reacts less to the introduction of
prescription drug cost-sharing than demand for less essential drugs.
In this paper, we add to the scarce but growing literature and investigate the issue of whether
demand for prescription drugs is sensitive to the price. We estimate the change in the propensity to
consume caused by a change in the price. All estimates of price sensitivity are uncovered within a
context where individuals may be forward-looking and within a market where private health
insurance that may cover part of or all costs related to prescription drugs exists. We believe that
these are the policy relevant parameters if one is interested in changing the current system
marginally.1 We use a rich register-based data set on a 20 % random sample of the Danish
population in the period 2000-2003. Contrary to many existing data sets, ours includes information
on therapeutic group, price, and out-of-pocket payment for every prescription drug purchase. These
data are augmented with socio-economic characteristics on a yearly basis. Besides superior data,
our main contributions are the following: 1) we use a regression kink design to overcome the
standard selection problem. The idea is very close to that of a regression discontinuity design (see
for example Lee and Lemieux (2009)) except that instead of a shift in levels, we exploit a shift in
slopes.2 Here we directly exploit that coinsurance payments decrease in a discontinuous fashion as
consumption (on a yearly basis) increases. In principle, there is no reason to think that individuals
just above a given kink point are different from individuals just below except for the fact that the
price faced by individuals below the kink point decreases less with total consumption than for
individuals above the kink point. Comparing the propensity to purchase for individuals who are just
above with that for individuals who are just below kink points then allows us to identify price
sensitivities. While we are not the first to acknowledge the existence of this type of identification,
see for example Guryan (2003) and Nielsen, Taber and Srensen (2010), we are, to the best of our
knowledge, among the very first to directly implement such a design. Recently, Card, Lee and Pei
(2009) have established conditions under which the regression kink design nonparametrically
identifies the local average response (Altonji and Matzkin (2005)) or, equivalently, the treatment
on the treated (Florens, Heckman, Meghir and Vytlacil (2008)). Card, Lee and Pei (2009) and
Nielsen, Taber and Srensen (2010) provide the only other direct implementation of a regression
kink design that we are aware of. 2) Health insurance is universally supplied by the Danish
government. This enables us to use a unifying identification strategy to consider estimates for the
entire Danish population as well as estimates for subgroups defined by socio-economic
characteristics. Using the same identification strategy for all subgroups makes it much easier to
compare estimates across subgroups. This is a major advantage compared to the existing studies
that either only have access to (or exploit) a relatively young (RAND HIE) or old (Contoyannis et
al. (2005), Chandra et al. (2010)) population or do not distinguish between subpopulations. We
demonstrate that estimates for relevant subgroups are of reasonable size and relate to each other in a
1
Of course, our estimates are uninformative about underlying price sensitivity in the absence of a private health
insurance market.
2
In particular the tax literature has devoted considerable attention to kinks (in budget sets); see Moffitt (1990) for an
early overview. Recently, these kinks have also been used as the basis for identification, for example to investigate the
responsiveness of labor supply to tax schemes; see Saez (2009) and Chetty et al. (2009). In fact, various non-linear
pricing schedules are often exploited to generate instruments in the economics literature; for recent examples see
Rothstein and Rouse (2007), Nielsen, Srensen and Taber (2008), and Dynarski, Gruber and Li (2009).
meaningful way, which clearly increases the credibility of the overall identification strategy. 3)
Finally, in addition to providing subpopulation specific estimates, we also distinguish between
types of drugs, something which is rarely possible simply because of lack of data.
Our results suggest that demand is inelastic; we find low average price responsiveness with a
corresponding price elasticity ranging from -0.09 to -0.25, thus comparable to the results from the
RAND HIE for the overall population. Individuals with lower education and income are, however,
more responsive to the price. The same is true for the elderly population, and our estimates are in
line with those of Chandra et al. (2010) and Contoyannis et al. (2005). Finally, essential drugs that
surely prevent deterioration of health and prolong life have, as expected, much lower associated
average price sensitivity than the complementary set of drugs.
The remaining part of the paper is organized as follows: Section 2 presents details of the Danish
market for outpatient prescription drugs and subsidy policies. Section 3 presents the conceptual
framework including a stylized economic model of drug purchase and links the model to parameters
of interest and identification strategy. Section 4 outlines the features of the available data. Section 5
gives the results from the empirical analysis, and Section 6 concludes.
2. The Danish Market for Outpatient Prescription Drugs
The Danish market for outpatient prescription drugs is highly regulated to secure correct handling
of as well as uniform prices on drugs across pharmacies. Pharmaceutical companies report
pharmacy purchase prices to the Danish Medicines Agency, who then announces retail prices.
These retail prices (along with a comprehensive list of information about the specific drugs
including substitutable drugs3) are made publicly available and registered online for five years.
Furthermore, changes in purchase prices must be reported by the pharmaceutical companies two
weeks in advance. Pharmacies must sell the cheapest substitute to prescription drugs unless the
prescribing doctor requires otherwise, or the patient specifically asks for another synonymous drug.4
Doctors do not have monetary incentives to prescribe more or less expensive drugs.
3
4
Substitutes are defined by having the same dose of the active substance as well as the same use (tablets, capsules etc).
Before 1997, the physician was required to write on the prescription if substitution was allowed.
Just as in the rest of OECD, consumption of prescription drugs has increased in a Danish context.
Figure 1 shows average level of consumption in Danish Crowns (DKK)5 in the period from 1995
2003 (in 1995 prices). On average, a Dane spent roughly DKK 1,750 on prescription drugs in 2003,
while the median was around DKK 400. In comparison, the average patient in treatment for
diabetes spent about DKK 4,400 on insulin and analogous products alone, the average patient in
treatment for excess gastric acid production spent about DKK 1,300 on relevant products, while
individuals in treatment with penicillin spent on average DKK 158.6
FIGURE 1
AVERAGE CONSUMPTION OF PRESCRIPTION DRUGS OVER TIME (DKK IN 1995)
1,900
1,800
1,700
1,600
1,500
1,400
1,300
1,200
1995
1996
1997
1998
1999
2000
2001
2002
2003
The subsidy scheme for adults 2000 - 2003
Subsidies were (and still are) based on the price of the products. Until June 25 2001, subsidies were
based on the average price of the two cheapest substitutes. This was designated the reference price.
5
100 corresponds to DKK 745 as of March 18 2009.

Insulin and analogous products identified by ATC code A10A, treatments for excess gastric acid production ATC
code A02A, penicillin ATC code J01C. See the Danish Medicines Agency for aggregate statistics,
www.medicinpriser.dk and p. 20 for a description of the ATC system.
6
After June 25 2001, this was replaced by a system where subsidies were based on the cheapest
product among substitutes. This was called the subsidy price. However, if one or more of the
substitutable products were sold within EU, the subsidy price would not be based on the Danish
price but on the average price within EU.7 Finally, products subject to parallel import received the
same subsidy as the original product sold in Denmark.
Starting in March 2000, individual level purchases of subsidized products (see section below for a
description) were entered into a central register, from which all pharmacies get their information,
and total costs were accumulated from the time of first purchase.8 Accumulated total costs (from
now on TC) are measured in reference/subsidy prices. Current TC is printed on receipts from drug
purchases and thus always available to a potential customer. Furthermore, the Danish Medicines
Agency provides a web page, www.medicinpriser.dk, where individuals can enter relevant
information and check the price they face for a potential product before going to the pharmacy. A
call to the pharmacy will yield the same information for those not connected to the internet.
TABLE 1
THE SUBSIDY SCHEME 2000 2003
Mar.-Dec. 2000
2001
2002
2003
DKK
DKK
DKK
DKK
0 % of TC
0 - 500
0 - 510
0 - 515
0 - 540
50 % of TC
500 - 1,200
510 - 1,230 515 - 1,240 540 - 1,300
75 % of TC 1,200 - 2,800 1,230 - 2,875 1,240 - 2,900 1,300 - 3,040
85 % of TC
2,800 +
2,875 +
2,900 +
3,040 +
Table 1 describes the subsidy scheme. If an individual has a TC below a given threshold prior to
purchasing and the price of buying a product brings TC above the threshold, the consumer will
receive the lower subsidy for the part of the price below the threshold and the higher subsidy for the
part of the price above the threshold. To give an example, consider an individual who considers
buying a product worth DKK 100. Assume that she has a TC of DKK 400 prior to the purchasing
decision. In this case, she is not subject to any subsidy should she decide to buy, and she will face
Prices in Greece, Spain, Portugal, and Luxemburg were excluded because income levels in these countries differed
substantially from that of Denmark.
8
Before March 2000, co-insurance rates were fixed at either 50 or 75 % depending on the type of product. An exception
was Insulin to treat diabetes, which was free of costs. See Skipper (2009) for an analysis of the introduction of the new
subsidy scheme.
the full price of DKK 100. Assume then instead that her TC prior to purchasing is DKK 450. Now
she will receive no subsidy for DKK 50 and a 50 % subsidy for the DKK 50 that brings her TC
account above the threshold point. In this alternative scenario, the price faced by the individual is
DKK 75.
A year after the first purchase of prescription drugs, TC is re-zeroed. The structure of the subsidies
forms the basis for identifying the effect of prices on demand for prescription drugs as described in
the following section.
Retail prices of most prescription drugs are subsidized by the government, leading to a general
subsidy, though some are only subsidized if the individual has a specific diagnosis or is officially
retired. This is called a conditional subsidy.
Apart from general and conditional subsidies, an individual can receive one-product, increased,
chronics, terminal, and municipality specific subsidies. In our data, we can identify the type of
subsidy individuals receive. One-product subsidies concern a specific type of product (and all its
substitutes) that is subject to neither a general nor a conditional subsidy. A general practitioner
makes the application on behalf of the patient, and the Danish Medicines Agency is the decisionmaking authority. If the subsidy is granted, all purchases of the given product will be added to TC
in the same manner as purchases of products with general or conditional subsidies. Typically, the
subsidy will be granted for life but may in certain cases be shorter (for example if the product is not
to be consumed over an extended period).
As pointed out above, general and conditional subsidies are based on the subsidy price. In very rare
cases, individuals are granted an increased subsidy based on a more expensive product. This only
occurs if the patient is allergic to a cheaper alternative or suffers from serious side-effects. The
application procedure is the same as for one-product subsidies, and, if granted, purchases are added
to TC. Again, the subsidy is typically granted for life.
Chronically ill individuals with a predicted yearly TC above a certain threshold (around DKK
18,000 in the 2000-2003 period) may receive full compensation for any purchases above this value.
The application procedure is the same as above, but subsidies are only for a five-year period. In
2001 and 2002, 9,084 and 8,141 people out of a population of about 5.5 million people received
chronics subsidies. Terminally ill patients who choose to spend the remaining time of their life at
home instead of being hospitalized do not pay for any drug purchases (inpatient prescription drug
consumption is free of charge as well). In 2001, 8,430 people received terminal subsidies, and in
2002 the number was 8,568. Finally, municipalities may choose to provide further subsidies. These
are typically granted to low-wealth, low-income retirees.
The private insurance market
As in most countries, there exists a private health insurance market as well. When we discuss our
identification strategy, we explicitly address the implications of this private option. There is one
player in the market for prescription drugs in Denmark, the company danmark. Crucial for our
study is that none of the policies of danmark change at the kink points described above. In fact,
none of danmarks policies change with yearly consumption of prescription drugs. Furthermore, it
is not possible to enroll if one has purchased any prescription drugs during the last 12 months.9
Relevant for our study, the company offers two types of policies: Type I covers all remaining outof-pocket expenditures related to products granted one of the government subsidies described
above. Type II covers half of the remaining out-of-pocket expenditures for products that receive a
government subsidy. A Type I membership costs around DKK 2,700 (in 2007) per year while Type
II membership costs about DKK 1,000.
We are unable to directly observe and merge membership status of danmark with our random
sample below. We do have access to a smaller representative survey of the Danish adult population
(Health and Sickness Survey, SUSY 2000) with roughly 4,200 respondents. This survey holds
information on health care utilization including membership status of danmark.10 In SUSY, 11 %
of the respondents hold a Type I insurance in danmark and 18 % hold a Type II insurance. 33 %
report that they are members of danmark but a few (2 %) do not remember their membership type
and a few (another 2 %) are sleeping members with the option to enroll in either Type I or II at a
later point in time. Since our estimation strategy detailed below is extremely data demanding,
however, it becomes futile to merge prescription drug purchases directly onto SUSY. Instead, we
use the information from SUSY to impute membership of danmark based on observable
9
Individuals aged 61 or above cannot enrol either.

SUSY is collected by the National Institute of Public Health. The data set is available from the Danish Data Archive.
10
characteristics. In Section Below, we perform sensitivity analysis with regards to (imputed)

membership, including type of membership.
To get a sense of which background variables are correlated with membership status, Table A1 in
Appendix A shows the results from a probit model of the propensity to take up insurance. We
investigate three sets of covariates; a) variables observed in both data sets, b) variables for which
we only observe a proxy in the survey, and c) variables only available in the survey but informative
about individuals who buy insurance. Our imputations are only based on variables in a).
Specification 1 only includes variables in a), while Specification 2 includes all variables. We see
that higher educated individuals without children and with above median income are more likely to
take up insurance. Members of danmark are also in better health and visit general practitioners
more often. This latter set of results could both be due to the fact that one must not have purchased
prescription drugs in the preceding 12 months to become a member and the possibility that
individuals who take up insurance generally invest more in their health.
3. Conceptual Framework
In order to emphasize the identifying assumptions behind our estimation strategy and to provide a
framework for interpreting our empirical results, this section firstly presents a stylized economic
model of drug purchase. The model is set within the regime described in Section 2 above, where the
crucial feature is that prescription drug subsidies vary with total consumption. We next discuss
parameters of interest and identification.
At each point in time, an individual experiences a risk of becoming ill and an associated need to buy
a given prescription drug. This risk is specific to the individual and may vary with characteristics,
observable as well as unobservable, and behavior. Therefore, at each point in time, the individual
must decide whether or not to purchase the drug. Since prescription drug purchase, by definition,
requires a prescription from a general practitioner, only individuals who experience medically
substantiated needs are able to do so. Individuals can decide to buy the drug in the amount
prescribed by the general practitioner or not; he cannot decide on a different amount or an entirely
different product. Because drug purchase is costly to the individual both in terms of foregone
consumption and time, not all individuals may choose to buy the drug, although it is expected to
10
alleviate pain and/or cure the disease (primary noncompliance), or individuals may take smaller
amounts than what is prescribed (secondary noncompliance), for example by taking what is known
as drug-holidays.
Consider a simple two-period model, 1,2. Assume for simplicity that there is only one kink in
the subsidy scheme outlined above and that both 1,2 lie within the same subsidy year (i.e.
before TC is re-zeroed). Other kinks can easily be incorporated. Assume that we are initially in a
neighborhood around the kink point. As hinted at above, this is the region that provides us with
identifying variation. If we are on the right hand side of the kink point, purchases in 1 will not
affect the price of potential purchases in 2. If we are on the left hand side, however, this is no
longer true. Formally let:

,
where the price without any subsidy is denoted , and is measured prior to the purchasing
decision.
Assume that utility is additively separable across time. Consider now the net value of purchasing
relative to not in period 1:
!
1 0

1 0 1
"#

! 1 $

where the left hand side is the net gain associated with purchasing, u indicates the contemporaneous
utility of DP1, is the contemporaneous costs associated with buying (monetary and otherwise),
is the discount factor, and A is the size of the subsidy at kink point A. ! is optimal future
consumption of some good worth the value of the reduction in the price of future drug purchase,
DP2. Here it is clear that the gain from purchasing today is the contemporaneous value of doing so
(1) plus the discounted expected value of a reduction in the price of a product purchased in period 2
(2).
11
Several things are important to note at this point. Firstly, if individuals discount heavily or do not
expect to buy tomorrow, (2) will be zero. Thus, we expect forward-looking individuals with chronic
conditions to react on the average rather than the marginal price, see Keeler, Newhouse, and Phelps
(1977), whereas individuals affected by an unexpected and temporary condition are more likely to
react on the marginal price. In our setting, a forward-looking individual with a chronic condition
knows that he will need treatment for a long period and that buying today will lower the price of the
product tomorrow. But secondly, and more importantly, if we are exactly at the kink point such that

, (2) will not contribute to the value of purchasing irrespective of type of disease and
degree of forward-looking behavior.
Price sensitivity may obviously vary with the product under consideration. It is likely that demand
for drugs prescribed for serious illnesses is less price sensitive than average demand. Also, the
nature of the symptoms associated with a disease may affect price sensitivity of demand. Finally,
individuals with different socio-economic characteristics may react differently. We investigate
some of these issues in our empirical analyses below.
3.1 Parameters of Interest and Identification Strategy
In principle, we are interested in uncovering the entire demand curve for a given prescription drug.
In other words, how does the propensity to buy a given product change with the price?
Unfortunately, the available data does not allow us to answer this question without strict parametric
assumptions. We can, however, non-parametrically uncover parts of the demand curve by
exploiting the structure of the subsidies described above.
Our identification strategy relies on the fact that the subsidy scheme introduces exogenous variation
in the price of a given product. Remember that we denote the price without any subsidy . Clearly,
given the subsidy scheme described above, the price P faced by the consumer in a neighborhood of
the first threshold value TCA is
(1)

'( )

&
$
'(
)
,+
1 $

) *
12
where TC is measured prior to the purchasing decision.
For a graphical representation, consider Figure 2. Here we depict the price faced by the consumer
for different values of TC starting below the first threshold value TCA.11 For low values of TC, the
price is constant, but as soon as buying the product will push TC above TCA, the price decreases
linearly with TC until the point where TC is exactly at TCA. Hereafter the price is constant (until
TCB).
Price
FIGURE 2
Accumulated Total Costs
When the price is constant, we obviously cannot hope to uncover any estimate of price sensitivity.
Similarly, when the price changes one-to-one with TC, we cannot distinguish the effect of higher
TC (and likely worse health) from the effect of lower price. What we can exploit, on the other hand,
is the shift in the slope of the P-TC curve at TCA, a so-called sharp regression kink design.12 It is
similar to a regression discontinuity design, except that instead of exploiting a shift in levels, we
11
Threshold values B, C, and D give rise to similar variation in the price, and the identification strategy is analogous.
We cannot exploit the shift in slopes furthest to the left because we are considering a range of products with different
prices, and the location of the first shift clearly depends on the price of the product without subsidy.
12
13
exploit a shift in slope. In such a setting, TC is called a forcing variable. In the following we will
describe the assumptions and mechanics behind the strategy.
Card, Lee and Pei (2009) formally outline the identifying assumptions behind the regression kink
design. We adopt their notation. Let first W be a set of (predetermined) unobserved random
variables with distribution function ,-, and let the distribution and density of TC conditional on
W be given by ./0|2 3|- and (/0|2 3|-, respectively. Finally, let the price P be a
deterministic function of TC, 4 ; let the purchase propensity be a function of price, total
costs and unobserved random variables, 5 1 6, , 7; and let predetermined
observed variables 8 97.13 X is determined before TC, which again is determined before P.
Assume the following:
(Regularity) 5 1 :;< 97 are real-valued function with continuous first derivatives.
(First stage) 4 is a known function that is everywhere continuous and is continuously
differentiable on ,
:;<
, , but lim/0A/0B 4C D lim/0E/0B 4C . In addition,
(/0|2 3
|- F 0 (G5 - H $ whereI
<,- F 0.
(Smooth density) ./0|2 3|- is twice continuously differentiable in tc at TCA for every w. That is,
JKLM|N OP|Q
JOP
is continuous in tc for all w.
Card, Lee and Pei (2009) show that these assumptions together imply that:
(a) 57 -| 3 is continuously differentiable in tc at TCA for all w.
(b)
RSTLMALMB
UVWXYZX[\|LM[LMB ]
UVWXYZX[\|LM[LMB ]
`RSTLMELMB
U^_
U^_
Ua^_
Ua^_
RSTLMALMB
`RSTLMELMB
U^_
U^_
b
Jcdecf
JOP
|
g
(c) 58 9h | 3 is continuously differentiable in tc at TCA for all x0.
Intuitively, with the above assumptions we can estimate our parameter of interest
Card, Lee and Pei (2009) note that X could in principle enter 6 directly. Leaving it out is without loss of
generality.
13
14
(2) b
Jcdecf
Jc
|
g
RSTLMALMB
UVWXYZX[\|LM[LMB ]
UVWXYZX[\|LM[LMB ]
`RSTLMELMB
U^_
U^_
Ua^_
Ua^_
RSTLMALMB
`RSTLMELMB
U^_
U^_
by comparing the slope of the propensity to consume with regard to TC for observations that are
just to the right of TCA with that of observations that are just to the left while properly correcting for
the deterministic shift in the relationship between P and TC. From this parameter, we can calculate
implied elasticities as well as predicted changes in amounts caused by changes in the price.
Specifically, let N be the number of potential buyers and i 5 1j be the quantity sold.
Clearly, then the percentage change in the propensity to buy caused by a percentage change in the
price just equals the classic price elasticity:
k
l5 1

l5 1j

li

l
5 1
l
5 1j l i
To give an example of our identification strategy, assume for simplicity that the propensity to
purchase a given prescription drug does not depend on health and thus in the absence of the subsidy
scheme would not correlate with TC14 but decreases linearly in the price of the product. In this case
the price variation from Figure 2 will translate into a propensity to purchase the product given in
Figure 3, where we see that the subsidy scheme introduces an exogenous change in the slope of the
propensity to consume at TCA.
14
In reality, the propensity to purchase prescription drugs may be correlated with TC. Individuals who are more often
ill, for example, are more likely to have a high TC.
15
Drug Purchase Propensity
FIGURE 3
0% Subsidy
0-50% Subsidy
50% Subsidy
Accumulated Total Costs
TCA
Discussion of the identifying assumptions and their implications
Apart from a set of regularity conditions, we rely on a shift in the P-TC curve at TCA along with the
assumption of smoothness of the first derivative of the density of TC conditional on W. As pointed
out by Card, Lee and Pei (2009), this latter assumption is the critical one and implies that agents
must not have full control of the forcing variable, TC. The discontinuous shift in the slope of the
relationship between P and TC arises immediately from the subsidy scheme as described above.
An important implication of the identifying assumptions is (c) above, which says that any
predetermined variable X should have a cumulative distribution function that is differentiable with
respect to TC. In other words, there must be no kink in the distribution of X. Card, Lee and Pei
(2009) stress that it is not enough to show that means of covariates are similar on both sides of the
kink point. Here, we need to consider the empirical distribution of X given TC. Of course, this is
trivially satisfied if the distribution of X does not vary with TC in a neighborhood around the kink
point. This would be true, for example, if individuals do not experience a shift in health status or the
propensity to take up private health insurance when their total costs increase slightly.
16
A second issue regarding identification concerns endogenous (or strategic) sorting with regard to
TC. In the tax literature, individuals sometimes bunch at tax kink points, see Chetty et al. (2009)
and Saez (2008). In that setting, bunching is optimal because higher income increases the marginal
tax rate. Therefore it might be optimal to refrain from supplying an extra hour of work. However,
we do not worry about bunching since it is suboptimal in our setting. Here, a higher level of
consumption weakly reduces the price faced by the consumer. We do, nonetheless, show the
distribution of observations around the kink point; see below.
Interpretation of parameter of interest
The parameter we uncover is clearly local. In fact, Card, Lee and Pei (2009) demonstrate that the
parameter uncovered by the regression kink design corresponds to both the treatment on the
treated parameter of Florens, Heckman, Meghir and Vytlacil (2008) and the local average
response of Altonji and Matzkin (2005). It can be interpreted as the expected price sensitivity
around TCA for individuals who buy drugs worth at least TCA or about DKK 500 (circa 70) in
2000 in a given 12-month period. Of course, this does not address the price sensitivity for
individuals who rarely (or never) buy prescription drugs. Similarly, the degree of price
responsiveness for the same individual may vary depending on the level of past consumption (i.e.
health status). Thus, even if all individuals who buy drugs worth a total of DKK 500 actually end up
buying drugs worth DKK TCB or about DKK 1,200 (circa 160) in 2000 in the same given 12month period, the estimated price sensitivity at TCA may differ from that at TCB.
A second issue is that we estimate price sensitivity in the presence of private health insurance. As
mentioned above, we do assume that individuals do not experience a shift in the propensity to take
up private health insurance when their total costs increase slightly. Still, private health insurance
matters for the interpretation of our results. Mechanically, because danmark pays a fixed share of
the costs to the patient, what the existence of private health insurance does is to shift the
relationship between P and TC; it effectively diminishes the extent to which P is reduced when TC
increases. In the extreme case where everybody subscribed to private health insurance and all costs
of prescriptions drugs were covered by this insurance, demand would not be sensitive to the price at
all. More generally, private health insurance reduces observed price sensitivity compared to a
17
regime with no private alternative. As described above, we investigate the sensitivity of our results
to the private insurance option via an imputation strategy; see Section 5.1. The same point can be
made for the additional government subsidies (chronics, terminal, and municipality specific) but as
we directly observe the receipt of these extra subsidies in our register data, we can directly
investigate if estimates for individuals who exclusively receive general subsidies are similar to
those for the overall population.
3.3 Estimation
Though our parameter of interest is non-parametrically identified, for efficiency reasons we impose
local parametric assumptions. We consider a small neighborhood around the threshold value and
estimate the propensity to purchase, using a simple probit model where our unit of observation is
whether a consumer buys prescription drugs during a given week.15 To demonstrate, consider the
following simple model where TC enters linearly in the index:
(3)
5 1| mWnh n n 1 F o # n 1 F o # ],
where TC is again the total cost variable, 1 F o # indicates whether a purchase was done just
above a given kink-point j, and 1 F o # is the interaction of total costs and the kink-
dummy. From the interaction term between the kink-dummy and TC in (2), we can calculate the
(estimated) difference in the propensity to purchase caused by TC crossing threshold value j.16 In
practice, we investigate whether higher order terms of TC should be included.
In particular, take the derivative of the above expression with respect to TC to get
l5W 1|1 F o #, ]
Wn n 1 F o #]p
l3
15
Whether we use a probit or a linear probability model does not change the conclusions from the analyses below. See
Table A5 in Appendix A.
16
The model is estimated under the restriction that there is no jump at the kink points. Yet the results are not sensitive
to this.
18
Since 1 F o # is dichotomous, we simply evaluate the above derivative for 1 F o # 1
and 1 F o # 0, and then take their difference:
l5W 1|1 F o # 1, ] l5W 1|1 F o # 0, ]

l3
l3
Wn n 1 F o #]p n p
This is our estimate of the numerator in (2). The denominator is the difference between
Jc
JOP
on each
side of the kink. This is immediately available from the price scheme presented above and clearly
depends on the kink in question.
A couple of practical problems arise within this framework: Firstly, our identification strategy
formally holds in a small neighborhood around the threshold points. In practice, we need to choose
a bandwidth. The problem with comparing observations that are far away from the threshold values
is, of course, that other factors beyond the difference in drug subsidy may drive the decision of drug
purchase. Individuals with lower TC are, for example, less likely to be ill in the first place. For
precisely this reason, we need observations close to a given value of TC. On the other hand, we
need a large number of observations. In general, a regression discontinuity design requires a large
number of observations; see for example Lee and Lemieux (2009). Intuitively, a regression kink
design is even more data demanding because we have to estimate not a shift in levels but a shift in
the slope. In the empirical analyses below, we investigate how sensitive our estimates are to the
choice of bandwidth.17
Secondly, if the price of the product under consideration (or equivalently TC) is too low relative to
the bandwidth for individuals on the left hand side of the kink, we run the risk that the price does
not depend on TC at all; we are on the leftmost flat part of the graph in Figure 1, see also (1) above.
If this is the case, we will not get a consistent estimate of our parameter of interest. In fact, if the
propensity to purchase is increasing in TC, we will most likely underestimate price sensitivity. In
the example in Figure 3 above, where we for expositional purposes assume that the propensity to
consume does not vary with underlying health, our estimate of the price sensitivity will be zero if
we use observations on the leftmost flat part as controls, although it is in fact negative. Conversely,
17
Appendix A shows the results from our main specification for a range of bandwidth choices.
19
if the price is too high, we run the risk that individuals cross the next kink point and receive an
even larger subsidy for part of the price. We therefore investigate the distribution of prices of
prescription drugs. If only a small share of prices are lower than our bandwidth, the problem that we
may partly identify of a part of the TC curve where the price does not depend on TC is of little
importance. In the same way, if only a small share of prices are higher than the difference between
the threshold points, the problem that individuals may cross the next threshold point is minor as
well. If we, for example, consider the 50 % threshold point, prices should be below DKK 700 in
2000 to avoid this problem. Finally, we investigate results for products where the price is
sufficiently large (larger than half the bandwidth) but also sufficiently small (smaller than the
difference between two neighboring kink points). In practice, excluding observations where the
price is too high or too low produces estimates that are essentially identical to the full-sample
results. Descriptive analyses of the distribution of prices and estimation results using this limited
sample are available on request.
Note that the fact that we need to estimate a shift in the slope of the propensity to purchase curve
precludes the use of falsification tests where a fake kink is investigated. Since the functional form
of the curve is unknown and potentially differs between real and fake kink points, we cannot use the
functional form specification from the actual kink points to investigate any fake kink points. Thus,
we cannot identify whether a significant shift in the slope of the curve at a random fake kink point
is just caused by misspecification.
Finally, in practice we can only meaningfully investigate price responsiveness in a neighborhood

around the lowest kink point TCA (and to some extent around TCB). This reflects both the paucity of
the data at higher kink points and smaller changes in subsidies.
4. Data and Descriptive Statistics
This section first describes the available data and discusses how we construct the dependent
variable as well as the forcing variable. We then present a range of descriptive statistics.
20
We use administrative data provided by Statistics Denmark. The data set contains information on a
representative sample of 20 % of all Danish individuals in the period from 2000-2003. For each
individual in the sample in this period, we know the complete history of prescription drug purchases
including date, price, amount of subsidy, type of subsidy, and type of drug. These data are
augmented with socio-economic information describing demographics, income, and education on a
yearly basis. Unfortunately, we do not know diagnoses, nor do we have information on unredeemed
prescription notes. Thus from the perspective of the econometrician, there is no difference between
a decision not to buy after having seen a doctor and having him prescribe a drug and the decision
not to go to a doctor in the first place. Finally, we do not know whether an individual has private
insurance. All estimates of price sensitivity are therefore estimated for potentially forward-looking
agents within a market where private health insurance that may cover part of or all costs related to
prescription drugs exists. As argued above, these are also the policy relevant estimates.
In the following, we discard observations for individuals who, at the time of purchase, were below
18 years of age. Most importantly, the subsidy scheme described above is only valid for adults.
Additionally, young individuals are perhaps more likely to have someone else pay for their
prescription drugs. Thus we would not measure their price sensitivity.
Table 2 first shows our variables related to prescription drug purchases. The Anatomical
Therapeutic Chemical (ATC) variable is a five level code for classification of drugs, which is
defined and maintained by WHO; see Table 3 for an example that explains the components of the
ATC-code.18,19 DDD measures the number of daily doses included on the prescription (given that
the drug is used for its primary purpose). TOS is the type of subsidy associated with the
prescription, see Section 2. SUB is the amount of subsidy received, while SP is the subsidy price
mentioned in Section 2. NAME is the brand or the name of the company that produced the drug. The
rest of the variables in Table 2 are self-explanatory.
18
Some prescription drugs are not assigned an ATC-code. This group often consists of so-called magistral medicinal
products, which are drugs produced at the pharmacy. In general, these products do not qualify for subsidy.
19
See also Appendix A, Table A2 for a list of the twenty most common (in terms of number of purchases) therapeutic
subgroups and Table A3 for the twenty largest therapeutic groups in terms of expenditure shares. Table A4 shows the
twenty most commonly sold products.
21
TABLE 2
VARIABLES DESCRIBING PRESCRIPTION DRUG PURCHASES
Variable
ID
NP
ATC
DDD
TOS
SUB
SP
NAME
EDP
TP
OP
Description
Individual identifier
Number of packages purchased
Active ingredient
Defined daily doses on prescription
Type of subsidy
Amount of subsidy received
Reference/subsidy price
Brand level name of drug
Exact date of purchase
Total price of drug
Out-of-pocket payment for drug
TABLE 3
EXAMPLE OF ATC-CODE
ATC-code A10BA02
A
Alimentary tract and metabolism
(1st level, main anatomical group)
A10
Drugs used to treat diabetes
(2nd level, therapeutic subgroup)
A10B
Oral blood glucose lowering drugs
(3rd level, pharmacological subgroup)
A10BA
Biguanides
(4th level, chemical subgroup)
A10BA02 Metformin
(5th level, chemical substance)
Table 4 next presents the list of variables describing socio-economic characteristics. All variables
are measured in the year prior to the purchasing decision. UNEMP specifies the fraction of
working-hours in a given year spent unemployed. INC is before-tax income of the individual, and
LINC is before-tax labor income. We use these variables to investigate price sensitivity for different
subgroups in the population.
22
TABLE 4
VARIABLES IN THE DATA SET
Variable
ID
AGE
NCHILD
UNEMP
INC
EDUC
GEN
LINC
Description
Individual identifier
Age
Number of children below the age of 18
Fraction of time spent unemployed
Yearly income
Highest completed education
Gender
Yearly labor income
The dependent variable
Our dependent variable is a dummy variable for prescription drug purchase in a given week. It takes
the value one if an individual purchases prescription drugs and zero otherwise. We can think of
three potential groups of individuals: Always-takers who buy a product regardless of the price,
never-takers who never buy a product, and compliers who buy a product if it is sufficiently
cheap, see Imbens and Angrist (1994). Since we do not know whether a prescription has been filled,
never-takers will likely constitute a sizeable fraction of the zeroes. On the other hand, some
individuals may not even go to the doctor in the first place because of price sensitivity. Remember
though that all the individuals we are exploiting for identification purposes have, by definition, at
some point during the last 12 months been to the doctor to pick up a prescription and made a
purchase. The fact that we do not observe degree of need is shared with the entire literature. It is,
however, innocuous in the sense that neither never-takers nor always-takers contribute to
identification of the change in the propensity to buy caused by a change in the price. All the
inclusion of these two groups does is to cause a parallel (downwards or upwards) shift in the
propensity to purchase drugs around the kink point. As such, they do not affect the slope of the
relationship between P and TC and have therefore no impact on the identification of the percentage
in the propensity to buy at a change in the price either. What will be affected, however, is the
estimate of the percentage change in the propensity to purchase at a percentage change in the price;
the associated implied elasticity . The reason is that this parameter is evaluated at the average
propensity to purchase; see the definition of the elasticity on page 14 above. As such, one can think
of the elasticity estimate as an upper bound of the elasticity among those with a prescription.
23
We consider all weeks in a year, so any seasonal differences in the propensity to purchase are
averaged out. Note that this means that individuals may appear more than once. We account for this
by clustering the standard errors at the individual level.20 For the purpose of constructing
elasticities, we consider the first purchase in a given week. This is done to avoid modeling the
decision to buy a basket of products at the same time. It is unproblematic as long as it is random
which product the pharmacist enters into the cash register as the first.
Constructing accumulated total costs, TC
We next need to construct the forcing variable, TC. As described in Section 2, TC is the sum of the
subsidy price associated with each purchase over the individuals subsidy year. We have
information on the date of purchase in the data as well as the subsidy price, so constructing TC
amounts to accumulating the subsidy price for each individual for all purchases starting from March
1 2000. 365 days after the first purchase TC is re-zeroed. The next TC year starts with the first
purchase after the re-zeroing.
4.1 Descriptive statistics
As pointed out above, our identifying assumptions imply that any predetermined variable X should
have a cumulative distribution function that is differentiable with respect to TC. I.e. there must be
no kink in the distribution of X. Figures 4-10 show the distribution of our predetermined covariates
around the 50 % subsidy kink. We consider number of children, labor income, unemployment, total
income, age, education and time in weeks until end of personal subsidy year. The latter is
potentially important to include in our conditioning set below because there is a non-decreasing
relationship between this variable and the size of TC and time until end of personal subsidy year is
likely to affect the individuals propensity to purchase since postponing consumption until the start
of the next personal subsidy year will lead to a dramatic decrease in the subsidy.21 Education is a
20
The results are not sensitive to clustering. Unfortunately, we are not able to incorporate individual level fixed effects
because only a very small fraction of individuals are observed twice with a TC within the bandwidth but on each side
of a given kink point.
21
Note that such stockpiling also precludes the use of this price/subsidy variation as the basis for uncovering price
sensitivity. In future work we plan to explore the stockpiling phenomenon. See also Skipper (2009) for evidence of
stockpiling in connection with the introduction of the current subsidy scheme.
24
dummy for more than 12 years of schooling. Income is measured in DKK and is discounted to year
2000. Again, all variables are measured in the year prior to the purchasing decision.
FIGURE 4a
DISTRIBUTION OF NUMBER OF CHILDREN AROUND 50 % SUBSIDY KINK
Number of Children
0.55
0.5
0.45
Number of Children
0.4
0.35
0.3
450
460
470
480
490
500
510
520
530
540
Averages are calculated within DKK 1 bins.
FIGURE 5a
DISTRIBUTION OF LABOR INCOME AROUND 50 % SUBSIDY KINK
Labor Income
150000
145000
140000
135000
130000
125000
Labor Income
120000
115000
110000
105000
100000
450
460
470
480
490
500
510
520
530
25
540
FIGURE 6a
DISTRIBUTION OF DEGREE OF UNEMPLOYMENT AROUND 50 % SUBSIDY KINK
Unemployment
45
43
41
39
37
35
Unemployment
33
31
29
27
25
450
460
470
480
490
500
510
520
530
540
FIGURE 7a
DISTRIBUTION OF INCOME AROUND 50 % SUBSIDY KINK
Income
225000
220000
215000
210000
Income
205000
200000
195000
190000
450
460
470
480
490
500
510
520
26
530
540
FIGURE 8a
DISTRIBUTION OF AGE AROUND 50 % SUBSIDY KINK
Age
55
54.5
54
53.5
53
52.5
Age
52
51.5
51
50.5
50
450
460
470
480
490
500
510
520
530
540
FIGURE 9a
DISTRIBUTION OF EDUCATION AROUND 50 % SUBSIDY KINK
Education
0.56
0.55
0.54
0.53
0.52
Educa tion
0.51
0.5
0.49
0.48
450
460
470
480
490
500
510
520
27
530
540
FIGURE 10a
DISTRIBUTION OF TIME IN WEEKS UNTIL END OF PERSONAL
SUBSIDY YEAR AROUND 50 % SUBSIDY KINK
Weeks left of subsidy year
33
32
31
30
29
weeksleft
28
27
26
25
24
450
460
470
480
490
500
510
520
530
540
For completeness and to illustrate the role of covariates in our estimations, Tables 5 and 6 show
differences in means for individuals in 2000 with a TC in the intervals between DKK 450-550 and
DKK 475-525. Results for other years and kinks are similar and available on request. We see that
some of the differences in means are statistically significant at the 5 % level when considering the
DKK 450-550 interval in Table 5. Since our sample is very large (about 100,000 yearly
observations in the 450-550 interval), this is expected. The differences themselves are very small.22
Considering the DKK 475-525 interval in Table 6 renders the differences in means close to zero.
22
The sign also varies from year to year.
28
TABLE 5
DESCRIPT IVE ST AT IST ICS ON SOCIO-ECONOMIC CHARACT ERIST ICS, YEAR 2000
450<TC<500
500<TC<550
NCHILD
LINC
UNEMP
INC
AGE
EDUC
WEEKS LEFT
Mean
Std. Dev.
N
0.34
0.78 50,270
100,108.57 145,713.70 50,270
30.08
125.14 50,270
185,754.94 183,015.61 50,270
56.82
18.03 50,270
0.45
0.50 50,270
39.32
10.37 50,270
Mean
Std. Dev.
N
t-statistic
0.32
0.77 45,753
-2.4
95,682.88 144,606.24 45,753
-4.7
30.42
127.13 45,753
0.4
183,172.47 151,168.10 45,753
-2.4
57.17
18.10 45,753
3.0
0.44
0.50 45,753
-3.1
38.71
10.47 45,753
-9.1
TABLE 6
DESCRIPT IVE ST AT IST ICS ON SOCIO-ECONOMIC CHARACT ERIST ICS, YEAR 2000
475<TC<525
475<TC<525
NCHILD
LINC
UNEMP
INC
AGE
EDUC
WEEKS LEFT
Mean
Std. Dev.
0.33
0.78
98,637.91 146,150.81
28.91
122.50
185,576.73 202,204.73
57.05
17.99
0.45
0.50
39.56
10.37
N
26,036
26,036
26,036
26,036
26,036
26,036
26,036
Mean
0.31
96,013.36
29.70
183,478.06
57.11
0.44
38.75
Std. Dev.
0.75
141,905.25
125.62
152,314.16
17.99
0.50
10.42
N
t-statistic
23,727
-2.3
23,727
-2.0
23,727
0.7
23,727
-1.3
23,727
0.4
23,727
-2.2
23,727
-8.7
Figure 11 shows the distribution of observations around the 50 % subsidy kink. The dots show the
average number of observations in DKK 1 intervals. As expected, there are no signs of bunching on
either side of the kink.
29
FIGURE 11a
DISTRIBUTION OF OBSERVATIONS AROUND 50 % SUBSIDY KINK
OBSERVATIONS
AT
50% SUBSIDY KINK
50000
45000
Observations
40000
35000
30000
25000
20000
450
460
470
480
490
500
510
520
530
540
550
TC
a
The average number of observations is calculated within bins of DKK 1.
5. Results
This section presents our estimation results. As described above we model the decision to purchase
prescription drugs in a given week, using the probit specification outlined in Section 3, but add year
dummies as well. Our main analyses consider the kink caused by the lowest subsidy (50 %). TC is
discounted to year 2000, using the consumer price index. To secure that we correctly capture the
functional form of the curve, we start out by including higher order terms of TC and test the model
down. In practice, all models include TC in levels, while higher order terms are insignificant.
Figure 12 shows the empirical relationship between the propensity to buy and TC around the lowest
kink point.23 The solid line shows the predicted values from OLS estimation of the model in Section
3, while the dots show the average purchase propensity in DKK 1 intervals. Important for our
23
It is the empirical equivalent to the right hand side kink in simple model example shown in Figure 3.
30
strategy, the figure indicates that there is a shift in the slope of the purchase propensity around the
50% subsidy kink point. This is the variation we are identifying price sensitivity off.
FIGURE 12a
PURCHASE PROPENSITY AT 50 % SUBSIDY KINK
Purchase Propensity
0,1275
0,1225
0,1175
0,1125
0,1075
450
460
470
480
490
500
510
520
530
540
550
TC
a
The average purchase propensity is calculated within bins of DKK 1. The solid line shows predicted values from OLS estimation of the model in
Section 3.
Table 8 presents the results from the formal analyses for the 50% subsidy kink point.24 As
mentioned above, all standard errors are clustered at the individual level. The upper part of the table
shows the results, using the full set of products. The estimates are small and negative and in line
with those from the existing literature. The size of the estimates does vary somewhat with the
bandwidth: the estimate using a bandwidth of DKK 50 yields an elasticity of -0.09; a 10 % increase
in the price decreases the propensity to buy with 0.9 %. This estimate is significant at the 10 %
level, while the estimated elasticity using a DKK 25 bandwidth is larger in size (-0.25) and
significant at the 5 % level. The lower part of the table shows results when including covariates in
the analyses. We condition on the variables shown in Tables 5 and 6 above: Number of children,
labor income, degree of unemployment throughout the year, income, age, an indicator for more than
12 years of education, and time in weeks until end of personal subsidy year. The size of the
24
Results where we for example include both first and second order terms of TC are available on request.
31
estimated elasticity using a DKK 25 bandwidth is reduced from -0.25 to -0.19 by this exercise, yet
the two estimates are not significantly different from each other. The result using a DKK 50
bandwidth is almost unchanged by the inclusion of covariates. Table A5 in Appendix A shows
these main results, both with and without conditioning set, for bandwidth choices ranging from
DKK 50 DKK 5 in DKK 5 intervals.
TABLE 8
TREATMENT EFFECTS AND IMPLIED ELASTICITIES, ALL PRODUCTS,
50 % SUBSIDY KINK, TC = DKK 500
(+/-)
Include covariates
(+/-)
50 DKK
25 DKK
Treatment effect
(*1,000)
-0.11
-0.34
50 DKK
25 DKK
-0.15
-0.26
S.E.
Implied Average Purchase
(*1,000) elasticity, price propensity
0.06
-0.09
87.96
0.12
0.17
85.47
0.12
-0.25
0.05
0.15
-0.12
-0.19
87.96
85.47
0.12
0.12
#
Observations
3,136,950
1,558,687
3,136,950
1,558,687
#
Individuals
263,391
171,564
263,391
171,564
T reatment effecs and implied elasticities at DKK 500 kink. S.E. clustered by Person ID. Results are not sensitive to clustering. Average price is
that paid by consumers. Italic estimates are significant at the 10 % level, bold estimates are significant at the 5 % level.
5.1 Sensitivity analysis
As discussed above, it is unlikely that all individuals react similarly to price variation. This
subsection investigates whether the estimated parameters for the 50 % subsidy threshold vary across
subpopulations and investigates price sensitivity at the 75 % threshold. Table 9 shows the results
where we only include individuals who exclusively receive general subsidies. That is, we exclude
individuals who receive any additional subsidies as outlined in Section 2. This exercise reduces
sample sizes with around 50 %. Significance is, not surprisingly, affected by this, but the estimates
are similar to those using the full set of products. Thus there is no evidence that individuals who
receive additional subsidies are more or less sensitive to the price of prescription drugs compared to
individuals who receive further subsidies.25
25
As mentioned above, in a future version we will investigate the sensitivity of our results to the exclusion of
individuals with private insurance. To some extent, the results of subgroup sensitivity analysis performed in this
subsection could be driven by differences across groups in the propensity to sign up for private health insurance. For
now, it is comforting to note that at least excluding individuals with access to additional public subsidies does not alter
the results.
32
TABLE 9
50 % SUBSIDY K INK, TC = DKK 500, GENERAL SUBSIDY ONLY
Treatment effect
S.E.
(*1,000)
50 DKK
-0.06
0.07
-0.06
96.44
0.09
25 DKK
-0.37
0.21
-0.29
93.38
0.12
(+/-)
Include covariates
(+/-)
50 DKK
25 DKK
-0.04
-0.13
0.03
0.10
-0.04
-0.10
96.44
93.38
0.09
0.12
#
Observations
1,550,985
612,289
1,550,985
612,289
#
Individuals
135,670
81,926
135,670
81,926
Secondly, we consider the consequences of excluding individuals with a high predicted propensity
to take up private insurance (membership of danmark). Specifically, we calculate the share with
private insurance from the SUSY data set in each of the cells made up by the cross-products of the
dummy following variables: income above median, more than 12 years of schooling, male, age less
than 30, age 30-64, age 65 or above, any children in the household. We define two exclusion
criteria: i) share with private insurance in cell higher than 0.5 and ii) share with private insurance in
cell higher than share in population (33 %). For precision purposes, we exclude cells with ten or
fewer observations in the SUSY data set. Table 10 presents the results.
TABLE 10
TREATMENT EFFECTS AND IMPLIED ELASTICITIES, LOW PROBABILITY OF MEMBERSHIP
Treatment effect S.E.
Implied Average Puchase
(*1,000)
Probability of private
insurance < 50 %
(+/-)
50 DKK
25 DKK
Include covariates
(+/-)
50 DKK
25 DKK
Probability of private
insurance < population prob.
(+/-)
50 DKK
25 DKK
Include covariates
(+/-)
50 DKK
25 DKK
#
Observations
#
Individuals
-0.14
-0.20
0.07
0.19
-0.10
-0.14
86.48
84.06
0.12
0.12
2,400,021
1,191,206
208,231
134,310
-0.17
-0.15
0.10
0.17
-0.13
-0.11
86.48
84.06
0.12
0.12
2,400,021
1,191,206
208,231
134,310
-0.24
-0.31
0.11
0.30
-0.18
-0.23
86.89
84.61
0.11
0.12
999,830
492,692
87,330
55,415
-0.27
-0.25
0.10
0.27
-0.20
-0.18
86.89
84.61
0.11
0.12
999,830
492,692
87,330
55,415
T reatment effecs and implied elasticities at DKK 500 kink. S.E. clustered by Person ID. Results are not sensitive to clustering.
Avg. P is average price paid by consumers. Italic estimates are significant at the 10 % level, bold estimates are significant at the 5 % level.
Probability of private insurance (membership of 'danmark') is 33 %.
33
When we only include individuals with a propensity to take up private insurance that is lower than
the share in the population, our estimated implied elasticities increase in size, though not
dramatically so, and remain statistically significant despite much reduced sample sizes.
Thirdly, we consider differences in price sensitivity by level of education and income in Tables 11
and 12, and the results are striking. We distinguish between high and low level of education (12
years or less education versus more than 12 years of education) and high and low income (less than
average income versus more than average income). Demand for prescription drugs for individuals
with lower levels of education is more responsive to the price than demand for individuals with
higher levels of education. Note that individuals with lower levels of education also pay a lower
average price. Similarly, demand for individuals with less than average income is more price
responsive than demand for individuals with higher than average income. There could be several
explanations for these patterns; apart from potential differences in preferences for health
investments and the demonstrated differences in the propensity to take up private insurance,
individuals with lower levels of income face tighter budget constraints, they could have less
information about the importance of taking a particular drug, or they could be treated differently by
doctors than individuals with higher socio-economic status. See for example Simeonova (2008).
TABLE 11
50 % SUBSIDY KINK, TC = DKK 500, BY EDUCATION
Treatment effect
S.E.
(*1,000)
#
Observations
#
Individuals
<12 yrs
(+/-)
50 DKK
25 DKK
-0.27
-0.41
0.09
0.25
-0.16
-0.24
79.13
76.90
0.13
0.13
1,529,311
758,851
136,858
89,783
(+/-)
50 DKK
25 DKK
0.03
-0.24
0.08
0.22
0.03
-0.22
98.42
95.68
0.10
0.11
1,607,639
799,836
127,729
82,229
50 DKK
25 DKK
-0.27
-0.22
0.08
0.22
-0.16
-0.13
79.13
76.90
0.13
0.13
1,529,311
758,851
136,858
89,783
50 DKK
25 DKK
-0.04
-0.29
0.07
0.20
-0.04
-0.26
98.42
95.68
0.10
0.11
1,607,639
799,836
127,729
82,229
>12 yrs
Include covariates
<12 yrs
(+/-)
>12 yrs
(+/-)
34
TABLE 12
50 % SUBSIDY KINK, TC = DKK 500, BY INCOME
Low Income
(+/-)
High Income
(+/-)
Include covariates
Low Income
(+/-)
High Income
(+/-)
Treatment effect
(*1,000)
S.E.
(*1,000)

elasticity, price propensity
#
Observations
50 DKK
25 DKK
-0.20
-0.51
0.08
0.22
50 DKK
25 DKK
0.01
0.00
50 DKK
25 DKK
50 DKK
25 DKK
#
Individuals
-0.12
-0.31
80.51
78.30
0.13
0.13
1,965,857
976,665
179,988
117,090
0.09
0.24
0.01
0.00
104.96
101.88
0.10
0.10
1,165,163
579,050
93,180
58,218
-0.09
-0.17
0.04
0.10
-0.05
-0.10
80.51
78.30
0.13
0.13
1,965,857
976,665
179,988
117,090
0.01
-0.02
0.04
0.12
0.01
-0.02
104.96
101.88
0.10
0.10
1,165,163
579,050
93,180
58,218
Table 13 shows the results for three age groups: individuals under the age of 30, individuals aged
31-64, and individuals aged 65 or above. Young individuals are literally insensitive to the price; the
estimates are close to zero and insignificant. Older individuals, on the other hand, are more sensitive
to the price of the product. One explanation for this pattern is simply life expectancy; if one does
not expect to live much longer, it may not pay off to invest much in health either; see the seminal
work by Grossman (1972) on health and Becker (1964) on human capital investments more
generally. Another explanation could be that the elderly population aged 65 or above also has lower
levels of income, though they also have higher accumulated wealth. Note that both the youngest
group (age below 30) and the oldest group (age 65 or above) are less likely to take up private
insurance than individuals aged 30-64; see Table A1.
35
TABLE 13
50 % SUBSIDY K INK, TC = DKK 500, BY AGE
Treatment effect S.E.
(*1,000)
< 30 years
(+/-)
30-64 years
(+/-)
65+ years
(+/-)
Include covariates
< 30 years
(+/-)
30-64 years
(+/-)
65+ years
(+/-)
#
Observations
#
Individuals
50 DKK
25 DKK
0.00
-0.02
0.14
0.39
0.00
-0.03
92.08
89.89
0.08
0.08
353,920
176,284
26,698
16,315
50 DKK
25 DKK
-0.07
-0.34
0.07
0.20
-0.07
-0.29
95.19
92.47
0.11
0.11
1,945,452
966,466
155,255
100,077
50 DKK
25 DKK
-0.27
-0.19
0.12
0.34
-0.13
-0.09
76.12
73.95
0.16
0.16
837,578
415,937
86,483
57,134
50 DKK
25 DKK
-0.05
-0.02
0.14
0.38
-0.06
-0.02
92.08
89.89
0.08
0.08
353,920
176,284
26,698
16,315
50 DKK
25 DKK
-0.12
-0.39
0.07
0.18
-0.11
-0.34
95.19
92.47
0.11
0.11
1,945,452
966,466
155,255
100,077
50 DKK
25 DKK
-0.30
-0.04
0.11
0.31
-0.14
-0.02
76.12
73.95
0.16
0.16
837,578
415,937
86,483
57,134
A fifth sensitivity analysis distinguishes between essential and other types of drugs (the
complement set). Essential drugs are defined as medications that prevent deterioration in health or
prolong life and would not likely be prescribed in the absence of a definitive diagnosis, Tamblyn et
al. (2001), page 422. See Table B2 in Appendix B for the list of drugs included in the essential
category. As expected, demand for essential drugs is less price responsive than demand for other
types of drugs; see Table 14. Note though that the complement set of drugs may also include drugs
that in some cases but not always fit the definition of essential drugs. One example is
antibiotics.
36
TABLE 14
TREATMENT EFFECTS AND IMPLIED ELASTICITIES
50 % SUBSIDY KINK, TC = DKK 500, BY DRUG TYPE
Treatment effect
S.E.
(*1,000)
Essential
(+/-)
#
Observations
#
Individuals
50 DKK
25 DKK
0.00
-0.12
0.03
0.09
-0.01
-0.35
101.55
99.29
0.03
0.03
3,136,950
1,558,687
263,391
171,564
50 DKK
25 DKK
-0.12
-0.22
0.05
0.14
-0.12
-0.21
82.39
79.84
0.08
0.08
3,136,950
1,558,687
263,391
171,564
50 DKK
25 DKK
-0.02
-0.09
0.03
0.08
-0.05
-0.27
101.55
99.29
0.03
0.03
3,136,950
1,558,687
263,391
171,564
50 DKK
25 DKK
-0.15
-0.17
0.02
0.13
-0.14
-0.16
82.39
79.84
0.08
0.08
3,136,950
1,558,687
263,391
171,564
Other
(+/-)
Include covariates
Essential
(+/-)
Other
(+/-)
We finally investigate the 75 % subsidy kink in Table 15. The estimated treatment effects are still
negative and slightly smaller in size. Only the result for the DKK 50 bandwidth is statistically
significant. Because of more limited sample sizes (the number of observations is reduced to around
40 % when we move from the 50 % subsidy kink to the 75 % subsidy kink) and a lower change in
the subsidy at the higher kink, we refrain from performing subgroup specific analyses.
TABLE 15
(+/-)
Include covariates
(+/-)
50 DKK
25 DKK
Treatment effect
(*1,000)
-0.43
-0.45
50 DKK
25 DKK
-0.50
-0.47
S.E.
0.20
60.76
0.17
-0.15
0.56
-0.16
59.76
0.17
0.19
0.53
-0.18
-0.17
60.76
59.76
0.17
0.17
#
Observations
1,289,790
639,732
1,289,790
639,732
#
Individuals
168,086
104,591
168,086
104,591
T reatment effecs and implied elasticities at DKK 1200 kink. S.E. clustered by Person ID. Results are not sensitive to clustering. Average price is the that
paid by consumers. Italic estimates are significant at the 10 % level, bold estimates are significant at the 5 % level.
Our results are not directly comparable to those from the RAND HIE (see Manning et al. (1987)
and Newhouse (1993)) since that study considered for the non-aged population total health care
utilization and not only prescription drugs. Our results, on the other hand, are local in the sense that
they are estimated around the 50 % subsidy kink point. Using a similarly aged population, our
37
results are, nonetheless, fairly close in size to those from the RAND HIE. Our study does suggest
that the elderly population is more responsive to the price than the non-elderly, and the results are in
line with the Canadian study by Contoyannis et al. (2005) who find moderate price elasticities (0.12 to -0.16) as well as with the US study by Chandra et al. (2010) (elasticities -0.08 to -0.15).26 As
suggested by Tamblyn et al. (2001) and also investigated by Chandra et al. (2010), we find that
demand for essential drugs is less sensitive to the price than less essential drugs.
6. Conclusion
We estimate price sensitivity of demand for prescription drugs, exploiting truly exogenous variation
in the price that stems from a kinked reimbursement scheme. Within a unifying framework, we are
able to address this question for different subpopulations and types of drugs. We find that demand is
indeed sensitive to the price, although estimated implied elasticities are small; the overall elasticity
ranges between -0.09 and -0.25 for individuals who have, so far, bought prescription drugs worth at
least DKK 500 ( 70) in a given 12-month period. There is important variation in which subgroups
are affected by the price of prescription drugs. Individuals with lower income and lower education
are, despite (or maybe because of) their lower average health capital, more sensitive to the price of a
product. The same is true for the elderly population. Thus, policy makers should be aware that
reductions in subsidies for these groups are likely to result in lower consumption and, presumably,
worse health outcomes. Along similar lines, lower consumption of prescription drugs may increase
the take-up of inpatient and outpatient cases; see for example Chandra et al. (2010) and Gaynor, Li,
and Vogt (2006) for evidence of this behavior. Finally, essential drugs that surely prevent
deterioration of health and keep patients alive have, as expected, much lower associated average
price sensitivity than other drugs.
26
Chandra et al. (2010) argue that the products with elastic demand are those for which consumers can easily substitute
into other treatments.
38
Literature
Altonji, J. G. and R. L: Matzkin (2005): Cross Section and Panel Data Estimators for Nonseparable
Models with Endogenous Regressors, Econometrica 73, 1053-1102.
Becker, G. S. (1964): Human Capital. New York: Columbia University Press.
Card, D., D. S. Lee and Z. Pei (2009), Quasi-Experimental Identification and Estimation in the
Regression Kink Design, Princeton University, WP # 553.
Chandra, A., J. Gruber, J. and R. McKnight (2010): Patient Cost - Sharing and Hospitalization
Offsets in the Elderly, American Economic Review 100, 193-213.
Chetty, R., J. N. Friedman, T. Olsen, and L. Pistaferri (2009): The Effect of Adjustment Costs and
Institutional Constraints on Labor Supply Elasticities: Evidence from Denmark. Mimeo, Harvard
University.
Contoyannis, P., J. Hurley, P. Grootendorst, S. Jeon and R. Tablyn (2005): Estimating the price
elasticity of expenditure for prescription drugs in the presence of non-linear price schedules: an
illustration from Quebec, Canada, Health Economics 14, 909-923.
Dynarski, S., J. Gruber and D. Li (2009): Cheaper by the dozen: using sibling discounts at catholic
schools to estimate the price elasticity of private school attendance. NBER WP # 15461.
Florens, J. P., J. J. Heckman, C. Meghir and E. J. Vytlacil (2008), Identification of Treatment

Effects Using Control Functions in Models with Continuous, Endogenous Treatment and
Heterogeneous Effects, Econometrica 76, 1191-1206.
Gaynor, M., J. Li and W. B. Vogt (2007): Is drug coverage a free lunch? Cross-price elasticities and
the design of prescription drug benefits, NBER WP #12758.
39
Goldman, D. P., G. F. Joyce, J. J. Escarce, J. E. Pace, M. D. Solomon, M. Laouri, P. B. Landsman,

and M. Teutsch (2004): Pharmacy Benefits and the Use of Drugs by the Chronically Ill, Journal
of the American Medical Association 291, 2344-2350.
Grossman, M. (1972): On the Concept of Health Capital and the Demand for Health, Journal of
Political Economy 80, 223-255.
Guryan, J. (2003): Does Money Matter: Regression-Discontinuity Estimates from Education

Finance Reform in Massachusetts, NBER WP # 8269.
Hahn, J., P. Todd and W. van der Klaauw (2001): Identification and Estimation of Treatment
Effects with a Regression-Discontinuity Design, Econometrica 69, 201-209.
Keeler, E. B., J. P. Newhouse, and C. E. Phelps (1977): Deductibles and the Demand for Medical
Care Services: the Theory of a Consumer Facing a Variable Price Schedule under Uncertainty,
Econometrica 45, 641-655.
Landsman, P. B., W. Yu, X. F. Liu, S. M. Teutsch and M. L. Baerger (2005): Impact of 3-Tier
Pharmacy Benefit Design and Increased Consumer Cost-sharing on Drug Utilization. The American
Journal of Managed Care 11, 621-628.
Lee, D. S. and T. Lemieux (forthcoming): Regression Discontinuity Designs in Economics, Journal

of Economic Literature.
Manning, W. G., J. P. Newhouse, N. Duan, E. B. Keeler, A. Leibowitz, and M. S. Marquis (1987):

Health Insurance and the Demand for Medical Care: Evidence from a Randomized Experiment. The
American Economic Review 77, 251-277.
Moffitt, Robert (1990): The Econometrics of Kinked Budget Constraints. Journal of Economic
Perspectives 4, 119-139.
40
Newhouse, J. (1993): Free for All: Lessons from the RAND Health Insurance Experiment.
Cambridge, MA. Harvard University Press.
Nielsen, H. S., C. Taber and T. Srensen (2010): Estimating the Effect of Student Aid on College
Enrollment: Evidence from a Government Grant Policy Reform. American Economic Journal:
Economic Policy 2, (May 2010): 185-215.
OECD (2009): Health at a Glance 2009: OECD Indicators.
Rothstein, J. and C. Rouse (2007): Constrained after College: Student Loans and Early Career
Occupational Choices, NBER WP #13117.
Saez, E. (2009): Do Tax Filers Bunch at Kink Points? Evidence, Elasticity Estimation, and Salience
Effects. Mimeo, University of California at Berkeley.
Simeonova, E. (2008): Doctors, Patients, and the Racial Mortality Gap: What are the Causes?
Mimeo, University of Stockholm.
Skipper, N. (2009): On Utilization and Stockpiling of Prescription Drugs when Co-payments

Increase: Heterogeneity Across Types of Drugs. Mimeo, Aarhus University.
Tamblyn, R., R. Laprise, J. A. Hanley et al. (2001): Adverse Events Associated with Prescription
Drug Cost-Sharing among Poor and Elderly Persons, Journal of the American Medical Association
285, 421-429.
41
Appendix A
TABLE A1
MARGINAL EFFECTS FROM PROBIT ESTIMATION OF MEMBERSHIP OF 'DANMARK'.
Specification 1
Specification 2
Marginal Effect
S.E.
Marginal Effect
S.E.
Income above median (0/1)
0.090
0.018
0.075
0.018
More than 12 years of schooling (0/1)
0.134
0.015
0.128
0.015
Male (0/1)
-0.117
0.015
-0.115
0.015
Age 30-64 (0/1)
0.068
0.020
0.066
0.020
Age 65 + (0/1)
-0.037
0.024
0.014
0.041
Children in household (0/1)
-0.039
0.017
-0.041
0.017
Unemployed during last three years (0/1)
0.018
0.044
Fully employed during last three years (0/1)
0.073
0.038
c) Additional variables in SUSY
Self-reported health 'perfect' (0/1)
0.036
0.016
a) Variables included in both register data and SUSY
b) Proxy variables in SUSY
Any visits to GP during past three months (0/1)
0.030
0.016
Regular use of pharmaceuticals (0/1)
-0.012
0.018
# Observations
2
Pseudo R
4,262
4,262
0.045
0.048
Marginal effects evaluated at the mean. T he excluded socio-economic group is retirement. Italic estimates are significant at the 10 %
level, bold estimates are significant at the 5 % level. Response rate in SUSY is 75 %.
42
THE 20 MOST
TABLE A2
FREQUENT LY USED THERAPEUT IC GROUPS
ATC-group
Count
Percent
Name
N05
2,538,410
9.3%
Psycholeptics
N02
2,515,002
9.2%
Analgesics
G03
1,883,471
6.9%
Sex hormones and modulators of the genital system
J01
1,730,247
6.3%
Antibacterial agents for systemic use
R03
1,720,373
6.3%
Obstructive airway disease agents
M01
1,387,090
5.1%
Anti-inflammatory and anti-rheumatic agents
C03
1,216,623
4.4%
Diuretics
N06
1,213,187
4.4%
Psycho analeptics
S01
1,056,132
3.9%
Ophthalmologics
C09
969,884
3.5%
Agents that exert an action on the renin-angiotensin system
A02
884,389
3.2%
Agents that exert an action on acid related disorders
C07
727,287
2.7%
Beta blocking agents
B01
720,792
2.6%
Anti-thrombotic drugs
A10
691,968
2.5%
Anti-diabetics
C08
654,489
2.4%
Calcium channel blockers
D07
603,136
2.2%
Corticosteroids for dermatological use
C01
537,727
2.0%
Cardiac therapy
A12
441,297
1.6%
Mineral supplements
C10
437,885
1.6%
Lipid modifying agents
R05
421,115
1.5%
Cough and cold preparations
Statistics Denmark. Purchases in period March 1st 2000-2003, age 18 and above, 20% sample.
As can be seen from Table A2, the type of drug most frequently used is psycholeptic drugs
(antipsychotics). The second most frequently purchased type of drugs belongs to the analgesics
category (pain relievers). This category covers products for severe pain (e.g. morphine) to over-thecounter products such as Panodil/Tylenol (mild pain relievers would qualify for conditional
subsidy). In third place are sex hormones and modulators of the genital system.
43
TABLE A3
THERAPEUT IC GROUPS BY EXPENDIT URES
b
ATC-group Total Percent
Name
R03
583.21
9.4%
Obstructive airway disease agents
N06
540.92
8.7%
Psycho analeptics
N02
485.86
7.9%
Analgesics
N05
416.05
6.7%
Psycholeptics
C09
384.01
6.2%
Agents that exert an action on the renin-angiotensin system
A02
363.43
5.9%
Agents that exert an action on acid related disorders
G03
347.65
5.6%
Sex hormones and modulators of the genital system
C10
294.26
4.8%
Lipid modifying agents
C08
285.55
4.6%
Calcium channel blockers
M01
220.28
3.6%
Anti-inflammatory and anti-rheumatic agents
N03
196.26
3.2%
Antiepileptics
A10
190.24
3.1%
Anti-diabetics
J01
181.32
2.9%
Antibacterial agents for systemic use
G04
144.35
2.3%
Urologicals
C07
127.90
2.1%
Beta blocking agents
C03
126.18
2.0%
Diuretics
S01
122.16
2.0%
Ophthalmologics
B01
94.63
1.5%
Anti-thrombotic drugs
N04
74.03
1.2%
Anti-Parkinson drugs
C01
72.45
1.2%
Cardiac therapy
Statistics Denmark. Purchases in period March 1st 2000-2003, age 18 and above, 20% sample.
b
In million DKK. 2000 prices.
Table A3 shows therapeutic groups by expenditure shares. Here, obstructive airway disease agents
(asthma medicine) dominate with psychoanaleptics (anti-depressants and ADHD drugs) and
analgesics in second and third place.
Finally, Table A4 shows the 20 most frequently sold products. Here, we exploit the full ATC-code
level. As can be seen, phenoxymethylpenicillin (penicillin) is the single most prescribed drug,
claiming about 2.5 % of total sales. The second most sold product is paracetamol, which is a pain
reliever. In third and fourth place are ibuprofen (used for anti-inflammatory and anti-rheumatic
purposes) and Tramadol (a pain-reliever).
44
THE 20 MOST
TABLE A4
SOLD PRODUCT S BY ATC-CODE
ATC
Count
Percent
2.6%
2.6%
1.9%
1.8%
1.8%
1.7%
1.5%
1.5%
1.4%
1.4%
1.4%
1.3%
1.3%
1.2%
1.1%
1.1%
1.0%
0.9%
0.9%
0.9%
J01CE02
701,131
N02BE01
698,499
M01AE01
530,563
N02AX02
503,580
C03AB01
495,914
C03CA01
461,611
B01AC06
406,307
G03AA10
402,766
A12BA01
387,657
N06AB04
375,156
N05BA01
371,741
N05CF01
362,669
C07AB02
354,240
C08CA01
322,811
N05BA04
305,946
M01AB05
294,313
G03CA03
280,244
R03BA02
254,655
R03AC02
251,607
R03AC03
249,014
Name
Phenoxymethylpenicillin
Paracetamol
Ibuprofen
Tramadol
Bendroflumethiazid and potassium
Furosemide
Acetylsalicylic acid
Gestodene and estrogen
Potassium chloride
Citalopram
Diazepam
Zopiclone
Metoprolol
Amlodipin
Oxazepam
Diclofenac
Estradiol
Budesonid
Salbutamol
Terbutaline
Statistics Denmark. Purchases in period March 1st 2000-2003, age 18 and above,
20% sample.
45
TABLE A5
(+/-)
Include all covariates

(+/-)
Include all covariates

OLS
(+/-)
50 DKK
45 DKK
40 DKK
35 DKK
30 DKK
25 DKK
20 DKK
15 DKK
10 DKK
5 DKK
Treatment effect
(*1,000)
-0.11
-0.13
-0.20
-0.34
-0.52
-0.34
-0.28
-0.52
-0.13
1.51
S.E.
(*1,000)
0.06
0.07
0.08
0.10
0.13
0.17
0.23
0.35
0.65
1.84
Implied
elasticity,
-0.09
-0.10
-0.15
-0.25
-0.38
-0.25
-0.20
-0.38
-0.09
1.05
50 DKK
45 DKK
40 DKK
35 DKK
30 DKK
25 DKK
20 DKK
15 DKK
10 DKK
5 DKK
-0.15
-0.18
-0.29
-0.40
-0.50
-0.26
-0.24
-0.57
-0.40
0.60
0.05
0.06
0.07
0.09
0.11
0.15
0.21
0.32
0.59
1.66
-0.12
-0.14
-0.22
-0.29
-0.36
-0.19
-0.17
-0.41
-0.28
0.42
50 DKK
45 DKK
40 DKK
35 DKK
30 DKK
25 DKK
20 DKK
15 DKK
10 DKK
5 DKK
-0.18
-0.24
-0.38
-0.50
-0.61
-0.37
-0.36
-0.75
-0.49
0.94
0.05
0.06
0.08
0.09
0.12
0.15
0.21
0.33
0.62
1.75
-0.13
-0.18
-0.28
-0.37
-0.44
-0.26
-0.26
-0.54
-0.35
0.65
T reatment effecs and implied elasticities at DKK 500 kink. S.E. clustered by Person ID.
Results are not sensitive to clustering. Average price is the that paid by consumers. Italic
estimates are significant at the 10 % level, bold estimates are significant at the 5 % level.
46
Appendix B
TABLE B1
ESSENTIAL DRUGS
Essential drugs:
Insulin, anticoagulants, angiotensin-converting enzyme inhibitors,

lipid-reducing medication, antihypertensives, furosemide, B-blockers,
antiarrhythmics, aspirin, antiviral medication, thyroid medication,
neuroleptics, antidepressants, anticonvulsants, antiparkinsonian drugs,
prednisone, -agonists, inhaled steroids, chloroquines, primaquines,
and cyclosporine.
47

Drugprices in Progress

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Drugprices in Progress

Încărcat de

Drepturi de autor:

Formate disponibile

Price Sensitivity of Demand for Prescription Drugs:

Exploiting a Regression Kink Design

Acknowledgements: We greatly appreciate valuable comments from Sren Leth-Petersen, Helena

2. The Danish Market for Outpatient Prescription Drugs

The subsidy scheme for adults 2000 - 2003

100 corresponds to DKK 745 as of March 18 2009.

The private insurance market

Individuals aged 61 or above cannot enrol either.

characteristics. In Section Below, we perform sensitivity analysis with regards to (imputed)

longer true. Formally let:

3.1 Parameters of Interest and Identification Strategy

where TC is measured prior to the purchasing decision.

Accumulated Total Costs

Assume the following:

is continuous in tc for all w.

(c) 58  9h | 3 is continuously differentiable in tc at TCA for all x0.

Drug Purchase Propensity

Accumulated Total Costs

Discussion of the identifying assumptions and their implications

Interpretation of parameter of interest

5 1|  mWnh  n   n 1 F  o #  n 1 F  o #  ],

l 5W 1|1 F  o # 1,  ] l 5W 1|1 F  o # 0,  ]

Finally, in practice we can only meaningfully investigate price responsiveness in a neighborhood

4. Data and Descriptive Statistics

The dependent variable

Constructing accumulated total costs, TC

4.1 Descriptive statistics

Averages are calculated within DKK 1 bins.

Averages are calculated within DKK 1 bins.

Averages are calculated within DKK 1 bins.

Averages are calculated within DKK 1 bins.

Averages are calculated within DKK 1 bins.

Averages are calculated within DKK 1 bins.

Averages are calculated within DKK 1 bins.

The sign also varies from year to year.

50% SUBSIDY KINK

The average number of observations is calculated within bins of DKK 1.

5.1 Sensitivity analysis

Implied Average Purchase

Becker, G. S. (1964): Human Capital. New York: Columbia University Press.

Florens, J. P., J. J. Heckman, C. Meghir and E. J. Vytlacil (2008), Identification of Treatment

Goldman, D. P., G. F. Joyce, J. J. Escarce, J. E. Pace, M. D. Solomon, M. Laouri, P. B. Landsman,

Guryan, J. (2003): Does Money Matter: Regression-Discontinuity Estimates from Education

Lee, D. S. and T. Lemieux (forthcoming): Regression Discontinuity Designs in Economics, Journal

Manning, W. G., J. P. Newhouse, N. Duan, E. B. Keeler, A. Leibowitz, and M. S. Marquis (1987):

OECD (2009): Health at a Glance 2009: OECD Indicators.

Skipper, N. (2009): On Utilization and Stockpiling of Prescription Drugs when Co-payments

Income above median (0/1)

More than 12 years of schooling (0/1)

Age 30-64 (0/1)

Children in household (0/1)

Unemployed during last three years (0/1)

Fully employed during last three years (0/1)

c) Additional variables in SUSY

Self-reported health 'perfect' (0/1)

a) Variables included in both register data and SUSY

b) Proxy variables in SUSY

Any visits to GP during past three months (0/1)

Regular use of pharmaceuticals (0/1)

Sex hormones and modulators of the genital system

Antibacterial agents for systemic use

Obstructive airway disease agents

Anti-inflammatory and anti-rheumatic agents

Agents that exert an action on the renin-angiotensin system

(c) 58 9h | 3 is continuously differentiable in tc at TCA for all x0.

5 1| mWnh n n 1 F o # n 1 F o # ],

l5W 1|1 F o # 1, ] l5W 1|1 F o # 0, ]